Framework integrations

PRME ships optional adapters for LangChain and LlamaIndex. Both use the public retrieval API and keep chat history in PRME’s immutable event log.

LangChain

Install the adapter dependency:

pip install "prme[langchain]"

Use PRME as a retriever or chat history:

from langchain_core.messages import AIMessage, HumanMessage
from prme.integrations.langchain import PRMEChatMessageHistory, PRMERetriever
 
retriever = PRMERetriever(
    directory="./memory",
    user_id="alice",
    top_k=5,
    token_budget=4096,
)
documents = retriever.invoke("What has Alice decided?")
 
history = PRMEChatMessageHistory(
    directory="./memory",
    user_id="alice",
    session_id="support-42",
)
history.add_messages([
    HumanMessage(content="Use dark mode."),
    AIMessage(content="I will remember that."),
])
history.clear()
assert history.messages == []
 
retriever.close()
history.close()

The history adapter preserves complete JSON-serializable LangChain messages, including tool-message identity, content blocks, names, and additional fields.

LlamaIndex

Install the adapter dependency:

pip install "prme[llamaindex]"

from llama_index.core.llms import ChatMessage
from prme.integrations.llamaindex import PRMEChatStore, PRMERetriever
 
retriever = PRMERetriever(directory="./memory", user_id="alice", top_k=5)
nodes = retriever.retrieve("What does Alice prefer?")
 
chat = PRMEChatStore(directory="./memory")
key = "alice:support-42"
chat.set_messages(key, [ChatMessage(role="user", content="Use dark mode.")])
chat.add_message(key, ChatMessage(role="assistant", content="Understood."))
removed = chat.delete_last_message(key)
assert removed is not None
 
retriever.close()
chat.close()

LlamaIndex keys use "user_id:session_id"; a key without a colon uses the session ID "default". set_messages, delete_messages, delete_message, and delete_last_message follow BaseChatStore behavior. get_keys() reports keys written through the current adapter instance because PRME does not expose a cross-owner key enumeration operation.

Retriever score floor

Both retrievers accept an optional min_score, the same inclusive floor as MemoryEngine.retrieve() (for example PRMERetriever(..., min_score=0.5)). Under rank fusion (PRME_SCORING__FUSION=rrf, the default) it applies to each result’s semantic_relevance, which is in the result metadata. With weighted scoring (PRME_SCORING__FUSION=weighted) it applies to the composite score. If vector search fails or the embedding model changed without a rebuild, no result has a cosine, so the floor is skipped instead of returning nothing, and every result’s metadata then carries min_score_skipped: True (RFC-0005 Section 7.2).

Append-only clear and delete

Clear, replacement, and deletion append versioned control events. The adapters apply those events when reading logical history, while the original messages remain available through PRME’s owner-scoped event and provenance APIs. Control events do not create graph nodes or enter vector and lexical retrieval. Reads page through the event stream and stop once the newest applicable clear marker makes older history irrelevant.

These operations change framework chat history. They do not archive memory nodes previously derived from a message. When an application intends to remove a memory from normal retrieval as well, resolve the message event through get_event_nodes() and apply the owner-scoped lifecycle operation explicitly.

History is isolated by owner, session, and exact PRME scope. Call close() when an adapter is no longer needed. Framework async fallbacks execute these sync methods in a worker thread.

Source: docs/FRAMEWORK-INTEGRATIONS.md  in PRME at 58caddb1.